AI EngineerGuide

input_transformations decoder for Claude and OpenAI

Paste a response, a stream or a 400 error body. Each input_transformations entry gets a plain-English meaning, whether it’s yours to fix, and the fix.

Runs in your browser. Nothing you paste is uploaded.

Test-mode tip from Anthropic's docs: fail only on an input_transformations entry whose reason is prefix_binding_mismatch. Entries with model_binding_mismatch or organization_binding_mismatch are expected after a switch.

9 reason codes, 4 error strings and 22 fixes, each quoted from an official Anthropic or OpenAI page (last checked 2026-11-20). Reason codes · Error messages

How to use it

  1. Paste an Anthropic or OpenAI response, an Anthropic SSE stream, a 400 error body, or just the input_transformations array.
  2. Optionally paste the request too, so each path (Claude) or id (OpenAI) resolves to the block it names.
  3. Fix the entries marked Yes. Model and account entries are expected after a switch.

How the decoder works

It finds input_transformations in whatever you paste and matches each entry’s type and reason against the official tables: Anthropic’s “What each input_transformations entry means” and OpenAI’s “Reasoning item removal across requests”. Unknown entries are shown raw, because Anthropic’s docs say to ignore an entry whose type or reason you don’t recognize.

Error bodies are matched on the exact phrases Anthropic says to match, in order: “The block is bound to a different conversation”, the beta-header sentence, block_binding: Extra inputs are not permitted, and the look-alike signature errors.

How we check this

Every outcome, reason code, error string and fix is data in reasoning_rules.json with a short verbatim quote, its source and the date we checked it. The examples on this page are the same fixtures the tests run. Anthropic pages were last re-read on 2026-11-20; OpenAI pages on 2026-11-19.

Unconfirmed in the docs, so shown as "unconfirmed" or "unknown": whether a string and a text-block array count as the same content on Claude; how the 400 error's position maps when system or same-role messages merge; OpenAI's capability order within a family; reasoning.context defaults on GPT-6.1 Sol and GPT-6.2; whether tools, text.format and reasoning.effort changes remove every reasoning item on GPT-6.2; and switch rules off the Claude API.

FAQ

What is input_transformations?

An array in the response that lists thinking blocks (Claude) or reasoning items (OpenAI) the API dropped or let through. On Claude it needs the thinking-binding-controls-2026-08-01 beta header and sits beside usage; on OpenAI it’s a top-level array of {id, action, reason, type}, in input order.

Why does the error's position differ from the entry's path?

The 400 error counts a role: "system" message, or one with the same role as the message before, as part of that earlier message, so the position it names can be lower than the index in your array. An input_transformations path is always the block’s place as you sent it.

Where is it in a stream?

In the message_start event’s message. After a mid-stream fallback, the final message_delta event carries the one that counts.